All posts
2 min readby Romiel Inolino

Microsoft's agent controls checks an agent's tool calls against your policy before they run, and returns allow, deny or transform

AI agentsagent governanceMicrosoft Foundryautomationapprovals

Microsoft wants business rules for agents to live in one policy service that every agent checks before it acts.

What happened

On November 18, Microsoft introduced agent controls, a managed Azure service now in public preview. It is built on two open-source pieces Microsoft published: Agent Hooks, which defines intervention points in an agent's run (startup, input, model calls, tool calls, output and shutdown), and the Agent Control Specification (ACS), a request and decision contract for evaluating what happens at each point.

At a hook, the agent host sends the proposed action or content to the service. Policies can combine Azure AI Content Safety classifiers, LLM judges and deterministic Rego rules, and the service returns allow, deny or transform. The host still enforces the verdict. Each evaluation also returns a receipt naming the policies and versions applied, so teams can trace an action back to the decision behind it.

Microsoft's worked example is a bank support agent: transfers over $10,000 go to human review at the pre tool call hook, an LLM judge flags financial hardship at input, and a PII classifier blocks or redacts personal data at output. The Learn docs give a simpler case, where a refund of $150 is denied against a $100 limit and $75 is allowed.

My take

Most of the agents I see in client systems carry their rules inside the system prompt: don't refund more than X, don't email anyone outside the account. Microsoft's own post calls a policy in a prompt a suggestion, and I agree. The model can ignore it, and nobody can prove afterwards which rule was in force.

The useful idea here is the split between the agent deciding how to do a task and a separate check deciding whether a specific action is allowed. Hard limits like refund caps, approved destinations and blocked admin tools belong in deterministic rules, and the judge models should only handle the cases that need judgment.

You don't need Azure to apply the pattern. In n8n or a custom agent, put a policy step in front of every tool that writes to the CRM, sends money or messages a customer, and store its decision next to the run ID. The receipts matter as much as the blocking, because they are what you show a client when they ask why the agent did something. The service is in preview and the post gives no pricing, so I would test it on one agent before planning around it.

More posts